Academic Writing skill for Codex

A strong paper states one contribution precisely, gives the evidence for it in an order a reviewer can check, claims no more than the evidence supports, and cites sources that exist and say what the paper says they say. Clarity is the goal; sounding academic is not.

Ground rules

• The author owns the science. Codex structures, edits, checks consistency and citations, and drafts connective prose from the author’s notes and results. It does not invent results, data, statistics, quotes or references.
• Citations must be verified. Never add a reference from memory. Every citation comes from the author’s library (.bib, Zotero export) or is looked up and confirmed (DOI resolves, title, authors and year match). Looking references up needs network; if it is off, list unverified citations for the author.
• Venue AI policies. Most journals and conferences require disclosure of AI assistance and prohibit listing AI as an author; some restrict AI-generated text or images. Ask for the venue and check its policy; keep a short note of what Codex did for the disclosure.

Before editing

1. Get the venue and its author guidelines: word or page limit, structure, citation style, abstract format, template (LaTeX class or Word), anonymization rules for double-blind.
2. Identify the article type: empirical study, systematic review, theory, methods, short communication, thesis chapter. Each has different expectations.
3. Check the matching reporting guideline (EQUATOR Network list): CONSORT (randomized trials), PRISMA 2020 (systematic reviews), STROBE (observational), ARRIVE 2.0 (animal research), SRQR or COREQ (qualitative), TRIPOD+AI (prediction models). Many journals require the checklist at submission.
4. Read the repo or folder layout: main.tex, sections, refs.bib, figures, a Makefile or latexmkrc, and any notes in AGENTS.md or a README about build commands.

Workflow

- [ ] 1. Write the contribution sentence and 3-5 key claims with their evidence
- [ ] 2. Outline sections so each claim has a home and an order
- [ ] 3. Revise section by section (introduction and discussion last)
- [ ] 4. Paragraph pass: one point per paragraph, topic sentence first
- [ ] 5. Claim-evidence pass: hedge or support every claim
- [ ] 6. Citation pass: existence, accuracy, consistency
- [ ] 7. Venue compliance: limits, template, checklist, anonymization
- [ ] 8. Build and proofread the PDF

1. Contribution first

Write, before revising prose:

Contribution: We show that <finding> in <setting>, using <approach>, which <implication>.
Claims:
C1 <claim> - evidence: Table 2, Fig. 3 (n=1,204; effect 0.31, 95% CI 0.22-0.40)
C2 <claim> - evidence: ablation, Sec. 5.2
C3 <claim> - evidence: case study, Appendix B (qualitative)

If the contribution needs more than one sentence, the paper may contain two papers or no clear contribution yet; tell the author.

2. Structure

Section	Job	Common failure
Title	State the finding or the question with key terms	Cute title with no searchable terms
Abstract	Context (1-2 sentences), gap, approach, main results with numbers, implication	No numbers; ends with “results are discussed”
Introduction	Funnel: problem, what is known, the specific gap, this paper’s contribution, roadmap	Long history before the gap; contribution hidden in the last paragraph without a clear statement
Related work	Position against the closest work by dimension, not a list of summaries	“X did A. Y did B.” with no comparison
Methods	Enough detail to reproduce; justify non-obvious choices	Missing preprocessing, exclusions, hyperparameters, software versions
Results	Report findings in the order of the claims, with effect sizes and uncertainty	Interpretation mixed in; p-values without effects
Discussion	Answer the question, relate to prior work, limitations, implications	Restating results; limitations as a token paragraph
Conclusion	The contribution and what it enables, briefly	New claims not supported earlier

Venue variants: ML and CS conferences often put contributions as a bulleted list at the end of the introduction; humanities papers develop an argument rather than IMRaD; follow the venue’s recent accepted papers.

3-4. Paragraphs and sentences

• One point per paragraph; the first sentence states it; the rest supports it; the last sentence links onward if needed.
• Old information before new: start sentences with what the reader already knows, end with the new point (where emphasis falls).
• Keep subject and verb close; avoid stacking nominalizations (“the performance of an evaluation of the optimization” becomes “we evaluated how well the optimizer…”).
• Use the same term for the same concept throughout. Synonyms for variety confuse readers into thinking there are two concepts. Define abbreviations once; limit them.
• Active voice and “we” are accepted in most fields now; follow the venue and the author.

5. Claims and hedging

Match the verb to the evidence:

Evidence	Appropriate verbs	Avoid
Randomized or controlled experiment, replicated	demonstrates, shows, causes (within scope)	proves
Single experiment or benchmark	shows, indicates	establishes, generalizes
Observational association	is associated with, predicts	causes, leads to, drives
Qualitative or case evidence	suggests, illustrates	shows that X is generally true
Speculation	may, we hypothesize, one possible explanation	(unhedged statements)

Flag every superlative (“first”, “novel”, “state of the art”, “significantly” without a test). “Significant” is reserved for statistical significance with the test named. Over-hedging (“may possibly suggest”) is also a problem: hedge once, in the right place.

6. Citations

Cross-check LaTeX citation keys against the bibliography:

grep -oh '\\[A-Za-z]*cite[A-Za-z]*\(\[[^]]*\]\)*{[^}]*}' *.tex sections/*.tex 2>/dev/null \
  | sed 's/.*{//; s/}$//' | tr ',' '\n' | sed 's/^ *//; s/ *$//' | sort -u > /tmp/cited.txt
grep -oh '^@[A-Za-z]*{[^,]*' *.bib | sed 's/.*{//' | sort -u > /tmp/defined.txt
comm -23 /tmp/cited.txt /tmp/defined.txt     # cited but not in .bib -> build will show [?]
comm -13 /tmp/cited.txt /tmp/defined.txt     # in .bib but never cited

Then check, for each citation that carries an argument:

• The source exists: DOI resolves (https://doi.org/<doi>) or the entry is in Crossref, PubMed, arXiv or the publisher’s site, with matching title, authors, year and venue.
• The source says what the sentence claims. Citing a paper for a claim it does not make is a common, serious error; when the author has the PDF, check the cited passage.
• Prefer primary sources over reviews for specific findings; reviews for overviews.
• Preprints later published: cite the published version when it exists.
• Style and fields consistent (biblatex with biber, or natbib with BibTeX); DOIs included if the venue wants them; title capitalization protected with braces in BibTeX ({DNA}, {Bayesian}).

Library management (Zotero, Better BibTeX, deduplication) belongs to $citation-management; surveying a body of literature belongs to $literature-review.

7. Venue compliance

- [ ] Word/page limit met in the compiled PDF, counting what the venue counts
- [ ] Required sections and statements: data availability, code availability, ethics
      approval, funding, conflicts of interest, author contributions (CRediT), AI disclosure
- [ ] Reporting checklist completed with page numbers
- [ ] Double-blind: no author names, acknowledgments, self-identifying links or "our prior work [12]"
- [ ] Figures readable in grayscale, fonts legible at print size, units on every axis
- [ ] Supplementary material referenced and consistent with the main text

8. Build and check

latexmk -pdf main.tex              # or -lualatex / -xelatex per the template
latexmk -c                         # clean auxiliary files
grep -n "Warning" main.log | grep -E "undefined|multiply defined|Overfull" | head -30
chktex -q main.tex | head -40      # optional LaTeX lint, if installed

Undefined references or citations in the log are blocking. Overfull boxes matter only when visible in the PDF.

Responding to reviewers

Produce a response letter that quotes each comment, answers it directly, states the change with the location in the revised manuscript, and stays courteous when disagreeing (give evidence, not tone). Mark every change in the manuscript as the venue requests (colored text or latexdiff old.tex new.tex > diff.tex).

**Reviewer 2, Comment 3:** The effect may be driven by the larger training set.
**Response:** We agree this needed testing. We retrained the baseline on the same 1.2M
examples (new Table 4); the gap narrows from 6.1 to 4.8 points but remains (p < 0.01,
paired bootstrap). Revised Sec. 5.3, lines 412-430.

Gotchas

• Fabricated or mismatched references are the fastest route to desk rejection and worse. Never add a reference that has not been verified.
• Introductions that review the whole field before stating the gap lose reviewers by page two.
• Numbers that differ between abstract, text, tables and figures. Check each number in the abstract against its table.
• Changing terminology mid-paper (model, system, framework, approach for the same thing).
• Limitations written defensively. State them plainly with their likely effect on the claims.

Verification

• The contribution sentence and claim list match the abstract and conclusion.
• Every claim has evidence in the paper or a verified citation; hedges match evidence strength.
• Citation cross-check shows no missing keys; argument-bearing citations confirmed to exist and support the claim, or listed for the author to confirm.
• The PDF builds without undefined references and meets the venue limits and checklist.
• Report what was changed, what needs the author’s decision, and anything unverified.